The increasing demand for accurate and scalable semantic enrichment of scientific publications has driven the need for hybrid frameworks that combine database-centric processing with the semantic power of Large Language Models (LLMs). We present the DETEXA+LLM Semantic Enrichment Workflow Builder, an extension of the DETEXA framework, which allows users to define, execute, and visualize hybrid text analysis workflows over relational backends. Our system supports declarative construction of pipelines that integrate regex-based pattern matching, classification, and selective LLM invocation via automatically generated Python UDFs and YeSQL queries. A visual web interface enables users to configure metadata enrichment workflows with minimal programming effort. We demonstrate the utility of our approach through three diverse and essential use cases for digital libraries: (i) classification of Data and Software Availability Statements (DAS), (ii) extraction of Protein Data Bank (PDB) codes, and (iii) mining of funding source mentions. Our results show that combining rule-based logic with targeted LLM calls improves both accuracy and efficiency, making the system suitable for large-scale scholarly metadata extraction tasks.
The growing need for scalable research resource discovery in digital libraries has motivated the development of frameworks that combine traditional processing with large language models (LLMs). Building upon DETEXA, a declarative and extensible SQL-based text analytics framework, we propose a key extension: the integration of LLMs through User-Defined Functions (UDFs). This architecture preserves the scalability and declarative nature of SQL workflows while selectively invoking LLMs for tasks that require deep semantic understanding. To balance the trade-off between efficiency and precision, we explore three processing alternatives: LLM-only pipeline, pattern-based approach, and a hybrid model with selective LLM invocation. As a running example, we focus on the extraction of Data and Software Availability Statements from research publications, demonstrating the benefits of hybrid semantic enrichment. Experimental results show that the proposed approach achieves significantly improved accuracy over pattern-based methods, while maintaining runtime performance compared to full LLM-based solutions. Our framework offers a path toward efficient, extensible, and semantically enhanced text analytics pipelines.
We study the WCAG compliancy and state of the metadata of PDF documents released under the Dutch Open Government Act (Woo). The results show that, in line with previous research on WCAG compliancy of PDF documents, only a fraction (0.2
OCR of historical handwritten documents is still a challenging task and an active research field due to the relatively low recognition accuracy achieved when processing manuscripts of different writing styles. In this work, we study the use of Large Language Models (LLMs) for correcting OCR in old Greek handwritten documents. We analyze two different old Greek datasets using a Deep Network based OCR along with several well-known and easy-to-use LLMs for correcting the output. Additionally, we generate synthetic erroneous texts and modify the LLM prompts to further investigate how LLMs perform in correcting noisy old Greek text. Experimental results show the potential of LLMs for OCR correction of old Greek handwritten documents, especially in cases where the recognition results are relatively poor.
One method to aid the understanding of a document corpus is to try constructing automatically a word/term/knowledge map for that corpus by analyzing the contents of the documents. Several methods have been proposed in the literature for this task. In this paper we investigate a novel method that is based on Association Rule Mining (ARM). ARM was proposed for databases, for structured data in general, as a method for data mining, e.g. for market basket analysis. Here we investigate its application over documents. In particular, we leverage association rule mining, through the Apriori algorithm, to find pairs of terms that co-occur in documents and their association. Each rule is characterized by its confidence and support. Then we map these rules to graphical elements. A key merit of the approach is that the user can interactively change the confidence and support threshold and obtain a different visualization. The evaluation over small datasets, up to datasets with 125.654 distinct words, showed that this approach is feasible and can produce maps that show the dominating words and connections. Source code: https://github.com/EfthimisM/AssociationMaps Video: https://youtu.be/eN9VrmmS6Ls.
When working with a corpus of documents (e.g. in the context of a digital library), various access services are offered such as browsing, keyword search, and faceted search. Recently, Retrieval Augmented Generation (RAG) approaches have been proposed that can leverage LLMs to offer Question Answering (QA) services while addressing the hallucination problem of LLMs. In this direction, this paper investigates an approach for offering QA over document corpora and related Knowledge Graphs that exploits LLMs, RAG, and RAG enhanced with information from Knowledge Graphs. To address the challenge of blackbox interaction, we present an interactive system called SemanticRAG. which enables users to ask questions, view the answer generated by each method, and obtain the provenance of each answer. We showcase the feasibility and value of this approach by deploying it over a corpus of scientific papers collected by the FAO UN for ecosystem restoration. Online Demo: https://demos.isl.ics.forth.gr/SemanticRAG/.
The increasing use of digital tools and materials in social sciences and humanities has generated a demand for research data management support. However, the provision of research data services for these disciplines remains underexplored. This study analyzes the implementation of research data management support for the social sciences and humanities from the perspective of academic libraries. Particularly, we explore existing research data services for these disciplines and investigate how research data management support for the social sciences and humanities could be improved. A participatory workshop with information specialists (n = 6) revealed three themes (research data services in practice, governance of the research data services, future of the research data services). Our findings contribute to understanding the current research data management support for the social sciences and humanities and increasing the maturity of research data services for these disciplines in academic libraries.
Modeling real-world events as structured graphs is essential for advancing research in information retrieval, digital history, and narrative analysis. In this paper, we propose StoryNetworks, a novel dataset that transforms short event texts into annotated event networks. We curated 5,204 events from the Wikipedia Current Events Portal spanning 2016 and 2017, and manually annotated 2,494 directed dependencies between them. By bridging unstructured textual data with graphbased event modeling, StoryNetworks offers a valuable resource for computational social science and digital humanities. In addition, as creating dependency graph from short texts is a challenging task, this dataset should be useful for designing new models in event evolution modeling, narrative structure analysis, and information diffusion to obtain better accuracy. The dataset is publicly available at https://github.com/ sumilab/dataset.
Author-assigned subject labels in Electronic Theses and Dissertations (ETDs) are often inconsistent, overly broad, or misaligned with the research focus. This hampers discovery, aggregation, and analysis, especially for interdisciplinary research. LLMs offer a scalable alternative for automated classification, but their labeling rationale is opaque and introduces systematic biases. This study compares subject labels generated by LLMs with human-assigned labels for over 9,000 ETDs across 21 academic categories to assess the disagreement. We evaluate multiple prompt-based and fine-tuned LLM configurations and analyze areas of agreement and disagreement to identify patterns of misclassification. LLMs achieve competitive performance overall but frequently misclassify theoretical or interdisciplinary texts, often due to overweighting lexical cues and disregarding context. We show such errors are not random but reflect structured semantic divergences from human interpretation. These findings suggest a need for hybrid frameworks that combine LLM scalability with human contextual judgment to improve subject labeling in academic repositories.
Several sustainable development goals, such as life on land, sustainable cities and communities, as well as good health and well-being, are heavily dependent on the migration of avian species from one end of the Earth to another. Machine learning tasks in this space rely on a supervised approach where a categorical output corresponding to a certain movement phenomenon is estimated leveraging geospatial and weather covariates. Although these covariates are sufficient for some tasks, they have limitations for some other tasks. We argue that some of these tasks can benefit from more expressive data, such as those generated by Large Language Models (LLMs). In this work therefore, we consider the task of augmenting spatio-temporal geospatial data with the output from an LLM for improved animal mobility prediction tasks. More specifically, first, we prompt an LLM and show it can be used to generate text data that can be used on its own for predicting animal movement phenomena, surpassing the performance of spatio-temporal geospatial data on some experiments. Second, we propose an algorithm to eliminate redundancies in LLM queries and reduce carbon footprints by finding coordinates in a mobility dataset that maximizes entropy. Third, we show that when these two data modalities are used together, the performance on the aforementioned prediction tasks is significantly higher than when each modality is used on its own. Finally, we propose a novel compositional attention framework to combine the two data modalities and select relevant features by alternating between geospatial covariates and LLM embeddings. Experiments on two tasks for reducing biodiversity loss via the forecast of migration states and another task for managing future global health risks via the one-health paradigm by estimating stop-over duration show that the proposed approach outperforms competing baselines.
The integration of language models into ecological workflows is opening new possibilities for automated species monitoring. Classification systems are especially relevant in this context, as the high volume of data generated by automated systems requires efficient tools to support expert curators. Multimodal approaches, which incorporate textual information alongside visual or acoustic data, have shown potential to improve classification performance and interpretability. However, for many insect taxa, structured and usable textual descriptions remain scarce or difficult to access. In this work, we present a tool for retrieving and merging textual information about moth species from official repositories and citable sources. The resulting descriptions can be used to enrich multimodal classification models across different taxonomic levels or to build structured databases for species comparison and discovery.
While citations are essential to scholarly communication, their role in assessing academic success has contributed to various forms of misuse. Citation cartels and citation padding are examples of such misuse, yielding irrelevant references and corrupting the academic record. Here, we introduce a tool that we developed and use internally for detecting problematic citation behaviors, called Cite Lens (sample code available on GitHub: https://github.com/MDPI-AG/citelens .), which analyzes citations using vector (embedding) similarity. This tool can either detect misalignment between an article and its references (article–reference similarity), or the reference and the paragraph in which it is cited (context–reference similarity). We analyze the citation patterns across multiple publishers and topics and show the capability of this approach to detect problematic citations. This tool aims to support MDPI’s editorial screening and help prevent unethical or manipulative citation behavior.
Citation-based metrics dominate academic evaluation in scientometrics but often favor researchers from well-represented regions, reinforcing systemic biases in scientific recognition. To address this challenge, we introduce a fairness-aware influence maximization framework for scientometrics, identifying top-k influential authors in citation networks by considering both the number of distinct citing authors and the geographic diversity of these authors. We introduce a multi-task learning model, Fair2Cite, that jointly learns embeddings of influencers and followers, accurately estimating influence probabilities while capturing latent behavioral traits. These embeddings help in the construction of a bipartite graph, enabling the selection of fair and impactful influencers via the Independent Cascade model. Experiments on scientometric dataset demonstrate that our method outperforms state-of-the-art baselines in influence spread and achieves a more balanced citation distribution across global regions. Our framework promotes an inclusive academic ecosystem by recognizing diverse contributions beyond traditional metrics.
Research findings in ecology have the potential to drive evidence-based actions that could reverse biodiversity decline, inspire nature-based solutions to climate change and enhance restoration of severely degraded waters and lands. However, publishing findings in peer-reviewed papers alone is not sufficient to turn ecological research into action, as evidenced by the burgeoning field of translational ecology. Scholarly literature remains inaccessible to many conservation and restoration practitioners. While the open access publishing movement has increased the availability of research, the knowledge is still poorly indexed and unstructured, leading to inadequate findability. We present a solution to these challenges in the form of the Ecolink Model (ELM) – an open-source schema for creating knowledge graphs that describe environmental variables, ecological processes and the relationships between them. Drawing on core concepts from ecological modeling and advances in biomedical knowledge synthesis, we outline a model written in LinkML – a domain-agnostic data modeling language – that captures the relationships at the heart of complex systems, thereby providing a structure for knowledge graphs. ELM establishes a consistent and reusable format that enables the discovery of new connections and presents knowledge in an easily searchable, intuitive way. Knowledge graphs that are constructed using ELM have the potential to enable restoration and conservation practitioners to easily access relevant research findings, to unveil new insights using graph data science techniques and drive an AI interface to provide plain-language access to ecological knowledge as described in the graph.
This study examines the evolution of coastal research from 1970 to 2023 using bibliometric analysis of Scopus data. We analyzed publications with "coastal areas" or "littoral" in their titles or abstracts, focusing on trends in research themes. By combining traditional topic modeling with Large Language Models (LLMs), we identified key research domains and tracked their evolution. Our findings reveal a significant increase in publications, particularly since the 2000s. The field has become more interdisciplinary, integrating social sciences and humanities to address complex coastal challenges. This analysis underscores the value of scientometrics in informing policy and research agendas, supporting the development of sustainable coastal management strategies.
As digital competence becomes a core priority for lifelong learning in Europe, the need to teach AI literacy grows increasingly urgent. To address this, we present IILAP, a browser-based Interactive Information Literacy Assessment Platform, designed to support the teaching and assessment of critical reading of AI-generated content. IILAP includes a Teacher Tool for curating chatbot QA datasets with truth labels and sources, and a Student Interface that provides responses enriched with citations and trust indicators. The system logs interaction data-such as time on task, source clicks, verification attempts, and error detection-to help educators identify gaps in students' critical reading skills. After each session, automated Excel reports summarize these measures for easy assessment. Developed through an initial user study, IILAP enables classroom deployment of controlled chatbot interactions and provides structured analytics aligned with indicators of critical thinking. This demo showcases how the system bridges user behavior and educational evaluation to fostering AI literacy in education.
In Dutch parliamentary debates, over 95
Search engines and LLMs are increasingly being used in learning contexts to find and access learning resources. While conventional ranking mechanisms in general-purpose search engines are based on topical relevance, in learning contexts, pedagogical suitability plays a crucial role in addressing learners' information needs, i.e. how well a resource supports a learner in expanding their knowledge within a specific context. This paper conducts an empirical study, investigating how search engine rankings and LLM rankings compare to those of human experts and learners to determine which ranking approach best supports learning. Using statistical methods, we analyze agreement across rankings collected from seven experts, 60 learners, and five LLMs over four topics. Results show that LLM rankings align more closely with expert judgments than with search engines or learners. Both experts and LLMs exhibit moderate internal agreement but differ notably from search engine rankings, indicating that conventional search engines are not optimized for pedagogical effectiveness.
Scalable biodiversity monitoring remains a critical challenge for global conservation, particularly in ecologically rich but underrepresented regions with limited data infrastructure. The AMBER project (Automated Monitoring of Biodiversity using Edge and Remote Sensing) addresses this gap by integrating lightweight, compressed machine learning models with the AMI insect-monitoring system to enable on-device species identification. We focus on moth classification as a tractable use case and evaluate two end-to-end inference pipelines: a full-featured, server-based baseline and a compressed, edge-optimised alternative. To support field deployment on low-power devices, we apply quantisation, and model distillation techniques and evaluate trade-offs between full-featured server-based inference and resource-efficient edge deployment strategies. Our results show that compressed models retain strong classification performance while drastically reducing computation and bandwidth needs, enabling scalable, real-time monitoring in remote settings. This work lays the foundation for scalable, real-time ecological monitoring through trustworthy edge AI systems.
Co-citation linkages are commonly used to measure implicit relationships between documents. Because traditional co-citation techniques have the drawback of treating the strengths of co-citation linkages equivalent within a single citing document, the analyses of citation contexts have proposed as solutions. Although an analysis that uses embeddings is one of the most promising techniques, such analysis has only been applied to co-citation linkages that are actually observed in citing documents. This study proposes a technique that identifies potential co-citation linkages via context-aware citation network embeddings. A potential co-citation linkage refers to a relationship between two documents of which one document is directly cited by a citing document, whereas the other is not; the uncited document is recommended to be cited in the same citing document based on the context-aware citation network embeddings. This study empirically evaluates the search performance of the proposed methods using potential cocitation linkages. The experimental results revealed that best-performing method outperforms baselines that target only co-citation linkages observed in the citing documents. Furthermore, the results indicated that the proposed method can appropriately incorporate potential co-citation linkages and can obtain relevant documents that are not identified by traditional co-citation searches.
